We accept the threshold

The question “when is the singularity?” is badly posed, because there is no single door. What we can measure is not a door but a curve: the length of the job a model can carry from start to finish. On METR's measurement that length has doubled steadily since 2019 — and over the last two years the doubling rate itself has almost doubled.

Serious people looking at the same data spread across 2026 to 2047. That is not a shortage of information; it is a sign that forecasting here is structurally hard.

A 2026 interview study found the reason: the answer correlates strongly with which building the answerer sits in.

We reject the sentence “it will happen anyway.” Epoch AI's analysis shows scaling is physically possible through 2030: modelled separately, none of the four constraints — power, chip supply, data, the latency wall — forms a wall. The first binding constraint is electricity: a frontier training run in 2030 will want gigawatt-scale power.

But what is possible does not happen by itself. Gigawatt grids, lithography lines, decade-long research programmes — these are human decisions, not laws of nature.

The inevitability claim is comforting because it excuses whoever makes it from responsibility. A person who says “it will happen anyway” never has to account for what happened. We do not want that comfort.

Here is what is lopsided about the debate: the risk of accelerating is always counted, the cost of slowing down never is. Yet every year of delay is a year spent with diseases that have no cure. That price is not abstract — it is countable. Nobody is counting it.

Amodei's “compressed 21st century” argument is useful here: if powerful AI can compress 50–100 years of biological progress into 5–10, then a decade of delay means deferring decades of medical progress.

This is not a trump card against safety objections — it is a call for bookkeeping. If you are trading one risk against another, write down both. An account that lists only one side is not an account.

The hardest part of a manifesto is writing down the evidence that weakens its own thesis. Here is that evidence.

ARC-AGI is the benchmark family that tries to measure intelligence not as skill but as skill-acquisition efficiency. As of February 2026, a cross-generation analysis of 82 approaches shows the following.

A system that cannot be controlled is not powerful; it is merely dangerous. Interpretability, evaluation and threshold commitments are not the opposite of speed — they are what carries it, because the first serious accident bills all of us for a decade-long pause.

Three frontier labs converged on the same structure: define a capability threshold, pre-commit to evaluations that detect approaching it, pre-commit to a response if it is crossed. DeepMind's latest framework (v3.1, April 2026) went a step further: it defines separate protocols for models that could accelerate AI R&D to “destabilising” levels, and brings large-scale internal deployments into scope. The intelligence explosion is no longer an academic argument; it is a company's formal risk category.

But we do not trust these frameworks blindly. The logic of pre-commitment weakens when the commitment itself changes five times in five months. We support them — and we want them independently audited.

One of the most rigorous empirical papers in the field asked: can software alone produce an intelligence explosion? The answer depends on the elasticity of substitution between research compute and cognitive labour. Two economists, at MIT and Yale, built a 2014–2024 panel for OpenAI, DeepMind, Anthropic and DeepSeek and estimated two production functions.

The decisions in this transition are being taken in a handful of cities, in a handful of boardrooms. Everyone else was assigned the role of audience — and we do not accept that role.

The most concrete finding on why this is urgent: of 25 researchers interviewed, 17 expect systems with advanced coding and R&D capability to be increasingly held inside companies and never shown publicly. The most frequently raised concern was exactly that: concentration of power, and progress accelerating behind closed doors.

If the most capable systems are never released, every external measurement is structurally behind. That does not make independent measurement pointless — it makes it the only remaining instrument of scrutiny.

Not spectating is a matter of contribution, not slogans: measuring, publishing data, translating, and holding the argument in our own language and our own public.

We want to multiply intelligence because people suffer, die, and live far below their capacity. That is the reason. We are not looking for another one.

There is a reason this article comes last. One of the most uncomfortable findings in the literature is the distinction between technological and social singularity: the second is the social discontinuity produced by the anticipated approach of the threshold. Even if the threshold never arrives, the expectation is already real — it is shaping capital allocation, energy infrastructure, regulation and labour decisions today.

Which means this manifesto is itself an intervention. We know that, and it is why we treat showing the basis of every sentence as an obligation.

If one day the trajectory of this technology stops treating the human as the end, everything this manifesto defends becomes void. We write that sentence here so we can come back and read it.